英伟达发布Nemotron 3.5 Transcribe，这是一款开放权重模型，可转写40种语言、翻译31种语言，并能在最长一小时的录音中标注最多八名说话人。与它一同发布的还有Granary v2，这是一个免费数据集，包含48种语言、共463,062小时的语音。
Nvidia released Nemotron 3.5 Transcribe, an open-weight model that transcribes 40 languages, translates 31 and labels up to eight speakers in recordings of up to an hour. It came with Granary v2, a free dataset of 463,062 hours of speech in 48 languages.
英伟达周三表示，已发布Nemotron 3.5 Transcribe，这是一款免费、开放权重的AI模型，可将40种语言的语音转为文字。同一模型还能将其中31种语言与英语互译、标注说话人身份，并回答关于录音的问题。
Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight AI model that turns speech into text in 40 languages, the company said on Wednesday. The same model can translate 31 of those languages to or from English, label who is speaking and answer questions about a recording.
它可处理最长一小时的录音，并能区分最多八名说话人。开发者通过自然语言提示来控制它，例如要求生成干净或逐字的转写文本，或提供医学、产品术语让它重点识别。
It can handle recordings of up to an hour and tell apart up to eight speakers. Developers control it with plain-language prompts, for example asking for a clean or word-for-word transcript, or giving it medical or product terms to listen for.
该模型拥有340亿参数，但每一步只使用约40亿参数，这使其运行速度快、成本低。它可在单块英伟达芯片上运行，从小的Jetson开发板到数据中心GPU均可，其许可证允许商业使用。
The model has 34B parameters, but only about 4B are used for each step, which keeps it fast and cheap to run. It works on a single Nvidia chip, from small Jetson boards to data-centre GPUs, and its licence allows commercial use.
英伟达称，该模型在Hugging Face英语语音识别排行榜的开放模型中排名第一，平均词错误率为5.33%。它专为录制音频设计，而非实时转写。
Nvidia says it ranks first among open models on Hugging Face’s English speech-recognition leaderboard, with an average word error rate of 5.33%. It is built for recorded audio, not live transcription.
与此同时，英伟达发布了Granary v2，这是一个免费数据集，包含48种语言、共463,062小时的语音，而第一版仅涵盖25种语言。开发者需按各原始来源自身的许可证从相应来源下载音频。竞争对手对同样的服务收费：SpaceXAI的Grok转写每小时收费10美分。
Alongside it, Nvidia released Granary v2, a free dataset of 463,062 hours of speech across 48 languages, up from 25 in the first version. Developers download the audio from each original source under its own licence. Rivals charge for the same job: SpaceXAI’s Grok transcription costs 10 cents an hour.
为展示该模型的可调优性，英伟达用印度医生的医学语音对其进行了训练。在留出的测试片段上，这将错误率降低了45.8%，在一块GPU上耗时约14小时。
To show how the model can be tuned, Nvidia trained it on medical speech from Indian doctors. That cut errors by 45.8% on held-out test clips, and took about 14 hours on one GPU.
此次发布进一步扩充了英伟达不断壮大的开放模型系列。此前已有4月的Nemotron 3 Nano Omni和8月的Nemotron 3.5 Lightning，后者为新模型的语言部分提供支持。
The release adds to Nvidia’s growing family of open models. It follows Nemotron 3 Nano Omni in April and Nemotron 3.5 Lightning in August, which powers the new model’s language side.